The main features are

- automated dd/fio tests

- combinations of kernel/fs/fio job/nr_dd/dirty_thresh that is
  hopefully complete enough to catch common performance regressions

- compare scripts for quickly evaluating performance and locate regressions

- detailed logs and graphs for analyzing regressions

To try it out,

1) copy all files to

        /path/to/writeback-tests

2) basic configuration

        cp fat-config.sh $(hostname)-config.sh
        vi  $(hostname)-config.sh
        vi config.sh

Minimal configuration is to prepare at least one empty partition and
point DEVICES to it, create one empty mount point and point MNT to it.

3) add a hook at the end of rc.local:

        /path/to/writeback-tests/main-loop.sh

main-loop.sh will test one case on each fresh boot.  It will
automatically reboot the test box for each test cases until all done.

Each test run will save its log files to a unique directory

        /path/to/writeback-tests/$(hostname)/<path1>/<path2>/

and if that directory already exists, the test case will be skipped.

We may further do a queue based job submission/execution system,
however this silly loop works good enough for me now :-)